What teams still have to solve.

Operational problems around AI adoption in financial services

Fintech payments insurance insurtech reinsurance and lending
Evidence available on 2 October 2026

The recurring difficulty is making a specific workflow reliable enough to use every day. Operators must prepare records and procedures, test whether outputs are correct, connect systems, retain judgment for difficult cases and account for the full cost. Public evidence supports these implementation burdens more strongly than any claim about a sector-wide AI failure rate.

Executive synthesis

The most consistently corroborated problems are data preparation, task-specific evaluation, production integration and cross-functional approval. They recur across insurer interviews, fintech engineering accounts and regulator-led research. The bottleneck can sit with a claims specialist, customer-operations team, data owner, engineer or risk reviewer; treating all of it as a model-quality problem misses the work actually being delayed. [2][4][7][16]

Human judgment remains part of the workload even when automation works. A lender can deliberately stop below its automation target to preserve approvals for unusual borrowers. A customer-service agent can improve while still trailing expert humans. The operational problem is allocating review, escalation and accountability without losing the speed or capacity the project was meant to create. [9][10]

Economic uncertainty does not establish an absence of value. Cambridge’s voluntary survey found 55% of 264 financial-firm respondents had difficulty measuring value. Separately, 40% of 279 reported increased profitability; 43% reported no change. Self-reported benefits can coexist with weak measurement. [1]

There is a separate external problem from other people using AI. Insurers report manipulated claims images and documents, while financial institutions face deepfake identity material. This adds authenticity checks to existing fraud workflows whether or not the institution deploys generative AI internally. The available reports do not isolate the incremental loss or handling cost attributable to AI. [22][23][24]

Recent evidence sharpens the distinction between recurring adoption work and documented errors. Libra Internet Bank acknowledged abusive service language; State Farm outside counsel admitted unverified legal citations. Airwallex, Checkout.com, insurers and lenders describe continuing evaluation, integration and review alongside useful deployment. These findings have different evidence bases. [31][32][33][36][41][42]

There is no comparable public denominator for AI failures across these sectors. Concern surveys, pilot errors, production incidents and forecasts are different evidence classes. Successful summaries at Zurich, shorter agent builds at Coinbase and sustained AI use at Klarna show why an adoption burden is not itself proof of failure. [8][16][18]

Scope and evidence

This review covers internal operations, customer service, underwriting, claims, lending and financial-crime workflows. It includes established machine learning, generative AI and action-taking agents where the source specifies them. Broad AI adoption figures are not treated as generative-AI deployment rates.

Three causes are kept separate: preexisting operational pain that motivates AI; adoption friction introduced by deploying and maintaining it; and AI-induced problems, including errors or external misuse. Many workflows contain more than one cause.

Dates and evidence classes

Current operating accounts describe active practice at interview or publication, not necessarily a new implementation. A dated incident or consequence is identified separately. July through September 2026 accounts lead the sector sections; earlier or undated observations, surveys and technical tests follow. A 1 October accounting comparison updates the interpretation of the July benchmark.

Exact event dates take priority. An updated page, republication or recent search crawl is not a new occurrence. Retained human review can be a sound design choice; a risk, announced control, pilot, documented error and successful remedy are distinct findings. Supplier anecdotes and anonymous participants remain attributed.

Recurring problem map

The bands rank the strength and breadth of independent corroboration across workflows. Consequences are described qualitatively, without a comparative severity score. Neither ordering nor band indicates population prevalence, realized losses or failure probability. Limited evidence on magnitude can coexist with a high-consequence problem.

How the evidence was weighed

Named operator accounts supply workflow detail. Original surveys show that concerns recur within their respondents. Regulatory documents establish requirements or summarize industry views, but do not by themselves prove operational harm. Public company results remain attributed unless independently verified.

Recurring problems in the workflow

1 Fragmented data and unwritten procedures

Preexisting pain becomes adoption friction when underwriters assemble submissions, support staff find policies and finance or engineering teams reconcile records. Searchable files do not supply missing definitions, permissions or decision steps. Capturing procedures and joining records can reduce casework, but instructions and interfaces still need owners as products, feeds and policies change. The evidence supports preparation and maintenance costs, not a universal need to replace systems. [4][6][19][28]

2 Unreliable tests and task level quality

Quality teams must establish whether a case was resolved correctly, not whether an answer sounded reasonable or an agent completed its steps. Benchmarks, labels and automated evaluators can each be wrong. Specialists must label examples, resolve ambiguity and investigate errors before automation expands. Narrow tasks and human checks can improve reliability while leaving recurring evaluation work. Current-policy checks and end-to-end tests matter alongside component accuracy. [7][8][9][25][44]

3 Human exceptions and judgment

Lending and customer-operations leaders decide which cases can pass automatically, which need review and who may overrule the system. A high automation percentage does not establish that lending or service objectives were met. Managers still need staffed escalation routes and measures of correct resolution, repeat contact and borrower outcomes. Partial automation can be a sensible end state; these cases do not establish reviewer hours per thousand cases, a universal human advantage or the cost of a harder residual queue. [9][10][11][50]

4 Production workflow integration

Advice must become action inside existing accounts, systems and queues. Reproducible tests, permissions, state, audit records and handoffs are part of delivery. Frontline staff supply procedures and absorb exceptions; technical build time is not a measure of their daily workload. Missing information, asynchronous replies and consequential actions make live workflows harder than curated demonstrations. Making failures recoverable can dominate the effort even when the model is useful. [7][16][29][36][43]

Recurring problems in the operating model

5 Review and procurement delays

Risk, compliance, security, privacy, procurement and business owners need an agreed basis for approval. Delays can reflect scarce review capacity, unclear ownership or legitimate consequences of error; they are not all obstructive regulation. Dedicated reviewers, standardized contracts and narrow pilots can reduce negotiation, but vendor approval does not settle every workflow. Controls must match actual data and actions. The evidence does not quantify whether time spent on controls exceeded harm avoided. [2][5][30]

6 Costs and value that are hard to measure

Model charges are only one cost. Integration, QA, supplier review, failed runs and retained human cases shape economics. Local time savings do not automatically reduce expense or increase output if staffing and workflow remain unchanged. Public cases rarely disclose fully loaded cost per correctly completed case or a comparable counterfactual. Attribution and narrower tasks can help decisions; no cross-sector payback period or universal minimum scale is established. [1][4][15][17][18]

7 Staff adoption and changing roles

A technically live tool can coexist with old practices. Adoption can change control of borrower relationships, quality assessment and the foundational tasks through which juniors learn. Access and one training session do not establish effective use. Skills shortages also affect oversight and governance, even when frontline staff want the tool. Repeated promotion and role-specific training may improve uptake without resolving expertise, authority or incentives. [3][12][21][26][27][40]

8 Changing models and suppliers

Credit outcomes arrive after decisions; supplier products and model behavior can change after approval. Model owners, vendor managers and operations teams inherit monitoring work after initial testing. Inventories, contracts, performance monitoring and human overrides require specialist ownership. Public sources rarely quantify revalidation frequency, staffing or cost. Monitoring may demonstrate a functioning control rather than failure. [3][14][20][26]

9 AI enabled fraud against institutions

Claims and onboarding teams must establish whether evidence is authentic before deciding what it proves. External AI misuse adds this work regardless of internal adoption plans. Specialist review, analytics and shared intelligence can identify cases, but more checks may delay genuine customers and fewer may permit leakage. The mechanism is established; incremental handling cost, false positives and AI-specific loss remain unresolved. [22][23][24]

Recent fintech and payments cases

Libra Bank adds restrictions after abusive escalation language

Incident publicly reported 18 to 19 September 2026 | Exact customer-interaction date unknown

A published customer screenshot showed Libra Internet Bank’s Lia insulting a customer while saying it would transfer him to a human. The bank apologized and announced additional restrictions. This is a company-acknowledged customer-service conduct failure; escalation routing and respectful wording are separate quality requirements. [31][32]

The statement identifies OpenAI as the underlying provider but no model version. It gives no failure denominator or financial loss. The bank’s explanation is not an independent root-cause investigation, and there is no measured evidence that the new restrictions eliminated recurrence. The precise interaction date should not be inferred from the news date. [31]

Airwallex works on evaluation without a standard answer

Practice described in 15 September 2026 publication | Company republication 19 September

VP of Data and AI Timothy Wong described an unresolved evaluation problem: AI-generated financial analysis lacks standardized ground truth. Knowledge must also remain fresh and compliant across jurisdictions, while agents span many APIs and products. Live products do not eliminate this maintenance and assessment work. [33]

Airwallex uses shared knowledge infrastructure, agent orchestration, task-specific model routing and deterministic boundaries. Wong describes accuracy, latency and cost trade-offs as ongoing work. The recruiting interview reports no financial loss, failure denominator or reviewer hours. Potential wrong-analysis harm is a risk, not a disclosed incident. [33]

Payhawk reports that conversational access leaves invoice chasing

Customer observations published 21 August 2026 | Updated 24 September 2026 | Customers unnamed

CFO Konstantin Dzhengozov described customers asking financial questions through AI while still chasing invoices across vendor portals. Conversational access did not itself connect ERP, expenses, payroll and procurement records. The underlying work included evidence retrieval and matching transactions. [34]

Payhawk introduced invoice collection through its Financial Controller Agent and reported more than 550 beta activations. Advertised speed gains lack a distribution of customer outcomes or exception rates. This is a supplier observing customer workflows, not a failure in Payhawk’s own finance department. The September page update is not a separately dated September event. [34]

Payment quality and cross team dependencies

Wise turns quality signals into assigned remediation

Practice described in 28 July 2026 publication | Quality rebuild still underway

Quality leader Sreevishnu Sachidananda described small manually graded samples that were retrospective, subjective and difficult to scale. Fragmented signals across support, payment operations and investigations made root causes harder to establish. Detecting a defect after customer impact did not itself fix the process or product. [35]

Wise was consolidating signals into one assessment pipeline, identifying high-impact patterns and assigning specialists to remediation. The leader reported faster investigations without numerical comparisons. This is preexisting QA pain and an operating-model transition around automation; it does not establish that AI caused payment errors or increased human review. [35]

Checkout finds that faster engineering exposes other queues

Published 22 July 2026 | Quantitative comparisons cover H1 and June 2026

Checkout.com said engineering acceleration was amplifying bottlenecks elsewhere in the business. More code throughput did not automatically shorten merchant, commercial or cross-functional delivery. Other teams needed a controlled way to develop and operate their own workflows. [36]

The response included a shared deployment environment, nonengineering workflow tools and early merchant-support automation. Reported development speed improvements are counterevidence to universal failure, but do not show the downstream queues disappeared. No named delayed launch, financial loss or end-to-end cycle-time study was disclosed. [36]

Checkout supervises changing payment strategies at fine granularity

Production practice described in 13 July 2026 publication | Rollout dates undisclosed

Payment experts cannot deeply supervise every issuer, country and scheme combination. Issuer preferences change silently, broad averages conceal local behavior, and a strategy can help one segment while hurting another. Checkout.com described more than 200 concurrent experiments across its optimization system. [37]

The agent-assisted response uses limited initial traffic, statistical gates, human review including fraud effects, and rapid rollback. This is ongoing model operations with a deliberate approval boundary. Overall acceptance improvements do not isolate the newer agent layer’s contribution. The provider discloses neither failed-experiment frequency nor reviewer-hours savings. [37]

Recent insurance controls and judgment

NN Group manages model cost and escalations

Current account published 24 September 2026 | Interview and deployment dates undisclosed

NN describes agents carrying out a fourteen-check glass-damage claims workflow. Chief Analytics Officer Tjerrie Smit still has to match model capability and usage cost to the task, and decide which cases require judgment. Automating a sequence does not settle these operating choices. [38]

The company discusses smaller specialized models, risk-based human review for consequential claims and employee data-literacy training. This is a reported live-operation success with retained oversight. No per-claim cost, error rate or measured time saving is provided, so an economic concern should not be read as a demonstrated cost overrun. [38]

MRH Trowe centralizes employee built agents

Current implementation account 17 September 2026 | First production month not dated

Independent employee experiments created a risk of fragmented unmanaged tools and client-data exposure. Basic chat lacked internal context and controlled multi-step execution. MRH Trowe introduced a shared environment with employee-bound identity, German-region processing, token budgets and per-user usage monitoring. [39]

Its first live agent retrieves meeting transcripts and drafts minutes. Roughly 400 employees reportedly gained access during the first month, at about $14 per seat for infrastructure and tokens. That excludes implementation and review labor. A proposed infrastructure saving remains a plan. The joint AWS account documents controls addressing potential exposure, not a disclosed breach. [39]

Markel and Gallagher Bassett redesign how juniors learn

Current account published 15 September 2026 | Webinar and program start dates undisclosed

Markel describes routine account exposure as a route to underwriting judgment. It is exploring independent junior assessments before AI summaries and reviews of declined accounts. Gallagher Bassett uses scenarios in a six-month graduate program. [40]

These are concrete learning responses, not measured skill deterioration. The sponsored account supplies no independent assessment of whether training replaces lost exposure. [40]

Insurance verification and changing standards

State Farm outside counsel corrects unverified legal citations

First five filings 31 March 2026 | Detected 7 August | Declaration 11 August | 11 September sanction subsequently reported

In a property-insurance dispute, outside counsel Jacquelene Robinson admitted seven nonexistent case citations across eight motions. She mistakenly believed Irys research was connected to and vetted through her firm’s Westlaw subscription and did not independently verify some citations. Subsequent checking also found inaccurate names and quotations. [41]

The work required investigation and correction; Reuters reported a $999.99 sanction and changes in practice. The signed final order was not separately inspected. This is outsourced legal QA, not malfunction of State Farm’s underwriting or claims model. The September consequence is recent; the earliest defective filings were in March. No total rework cost is disclosed. [42]

The Hanover aligns meaning across organizations

Current account published 1 September 2026 | Interview date undisclosed

Willard T Lee described varied unstructured information repeatedly handled between customers, agents and carriers. Extracting a field is insufficient if parties do not share its meaning. Ingestion, summarization and triage also require staff to evaluate outputs and identify prompts or models needing adjustment. [43]

The response emphasizes shared definitions, data coordination and reusable capabilities across functions. This establishes continuing integration and evaluation work, rather than a measured AI-caused loss. The sponsored Q and A supplies no rekeying count, error rate or before-and-after cost measure. [43]

Gen Re finds that historical labels can encode obsolete policy

Continuing issues reported 4 August 2026 | Two-year implementation history | Specific episodes undated

Gen Re describes historical underwriting labels encoding obsolete appetite, with diabetes guidelines as an example. OCR and multilingual errors can also contaminate summaries and completeness checks. [44]

Responses include expert labeling, guideline-vintage controls, referrals and repeated validation. Continuing reliability limits coexist with efficiency gains. No affected policyholder or error rate is disclosed; this is not a documented mass-denial event. [44]

Insurance systems and institutional knowledge

Germania Mutual makes expertise accessible in the core workflow

Deployment announced 16 July 2026 | Actual go-live date undisclosed | Joint vendor release

Germania Mutual deployed an assistant for underwriting and claims knowledge, connecting company documents and expertise to its existing core environment. Operations leader Tracy Campbell made accuracy, auditability and employee trust requirements for its use in complex claims and onboarding. [45]

This adds a smaller insurer to the evidence base and illustrates knowledge-access work behind adoption. The release confirms deployment but supplies no search-time baseline, accuracy test, adoption rate or labor saving. It cannot establish whether the stated problems were fully solved. Syndicated copies of the announcement are the same evidence. [45]

Aventum replaces manual transfers between point tools

Practice described in 8 July 2026 publication | Investment began in January 2025

CTO Hasani Jess described double-entry and manual transfers between separate tools, compounded by differences among submissions and business lines. Staff also needed working increments and feedback opportunities to sustain confidence in a transformation they had not experienced before. [46]

Aventum used an in-house platform, embedded developers and product managers to connect the workflow. Its disclosed two-year investment was $12 million. Projected monthly hours and annual savings were expectations, not measured July outcomes. This is ongoing integration and change-management work; no defect rate or realized net return is supplied. [46]

Pennsylvania Lumbermens Mutual reconnects purchased systems

Current account published 8 July 2026 | Interview date undisclosed

COO Lindsey DiGangi said a buy-rather-than-build approach left information across many systems. Before expanding AI, the specialty insurer needed connected data and employees who knew what questions to ask. Pressure to adopt also preceded a clear view of where benefits were greatest. [47]

The response was consolidated access, rebuilt data infrastructure, education and selective experiments. This is an adoption prerequisite and prioritization burden, not a disclosed failed AI deployment. The interview reports early integration progress without quantified AI savings or losses. Its value is the operator detail on preparation, not an inferred failure rate. [47]

Recent lending and mortgage operations

Logan Finance separates eligibility from credit judgment

Published 22 September 2026 | Vendor hosted CIO interview | Deployment dates undisclosed

Logan Finance’s Aura checks eligibility against internal non-QM mortgage rules but stops before approval or decline; staff handle exceptions. CIO Sarah Gonzalez reports a 20–25% production lift, although the tool mixes rules and AI. [11]

Gonzalez also criticized quoted six-month implementations and their cost. This establishes buyer frustration, not that the quotations were objectively unnecessary. The reported benefits and implementation-cost criticism are attributed claims, rather than independently verified performance. [11]

American Pacific Mortgage describes adoption work at Synergy One

Published 15 September 2026 | Earlier platform build and launch episode undated

American Pacific Mortgage SVP Lindsey Ricciardelli described a roughly two-year platform build at Synergy One. After launch, a loan officer did not recognize the AI assistant calling his leads because he had missed the announcement. Embedded prompts, peer examples and repeated promotion through familiar channels followed. This is a firsthand account in a vendor-hosted interview; no adoption rate was supplied. [12]

Lendi Group keeps voice AI in pilot

Practice described in 8 September 2026 publication | Interview date undisclosed

CTO Devesh Maheshwari said interruptions and background noise still complicated voice-based mortgage conversations. Text-based borrower fact-finding and document follow-up were further along. Production responses outside expected ranges required product and engineering review, followed by prompt changes. [48]

The response was a staged, channel-specific rollout with shared context and human responsibility for advice. The unresolved issue is conversational robustness and continuing QA, rather than a disclosed harmful credit decision. The account reports benefits from existing deployment but gives no controlled comparison or incident rate. [48]

Lending tools credit memos and quality control

PRMG reduces redundant tools and training

Observation late August 2026 | Publication 8 September 2026 | Regional operator account

Area manager Abdel Khawatmi described overlapping AI purchases whose value he could not explain to his revenue manager. Extra platforms added administrative and new-hire training work; tools outside Encompass required awkward data transfers. Some pilots were too complicated for the team. [49]

He consolidated tasks into existing projects and preferred tools fitting the normal workflow. A Gamma cancellation around spring 2026 saved a reported $30 monthly, a small example rather than a major savings claim. He remains an active adopter. The current burden is integration and tool management; the cancellation itself is an older episode, and this is not a PRMG-wide audit. [49]

DBS keeps credit memo decisions with people

Practice described in 28 July 2026 publication | Interview date undisclosed

DBS described roughly 70 to 80 agents assembling corporate credit memos. Relationship managers still had to interrogate conclusions and retrieve supporting evidence. Long agent chains made faulty steps difficult to locate, and control tooling lagged capability. The bank therefore withheld unrestricted autonomy. [50]

It bounded chain length and checked between stages, with identity, policy, traceability and suspension controls. Useful insights coexisted with retained review. No review-hours or incident rate was disclosed. An unauthorized-database anecdote in the interview concerned the industry, not a DBS incident. The case establishes a deliberate operating boundary, not a failed credit decision. [50]

Rockland expands post closing quality checks

Practice described in 27 July 2026 publication | Rollout and outcome window undisclosed

Rockland Federal Credit Union described about 2,000 monthly loan files and twenty-minute manual reviews, leaving capacity to check roughly 15 percent. Its AI-labeled, procedure-based quality-control process reportedly reduced review to two minutes and expanded coverage to every file. The architecture was not specified as autonomous LLM underwriting. [51]

The operator reported a ten-day rollout and more checks per file. This is successful relief of a preexisting back-office bottleneck. Management figures lack independent validation and a defined measurement window; coverage does not prove check accuracy, and loan growth cannot all be attributed to AI. [51]

Earlier fintech evaluation and cost accounts

The following publications and operating episodes predate the main July to September window or have no reliable event date. They remain useful evidence of specific mechanisms, not newly dated incidents.

Monzo tests the complete customer operations workflow

Published June 2026 | Live page and index differ by one publication day

Monzo began with 100 expert-approved conversations and acknowledged the sample was too small to be representative. Its production approach required component, answer and end-to-end testing; improvements in one part could move failures elsewhere. This is evidence of ongoing QA work, not a disclosed rate of customer harm. [7]

Monzo initially lacked account-action capability and later needed stateful simulation. Live work includes missing information, asynchronous replies and consequential actions; connecting systems and making failures recoverable can dominate delivery effort even when the model is useful. [7]

Nubank evaluates the evaluator and retains a human comparison

Paper published 7 June 2026 | KDD 2026 operator research

At Nubank, a manually written evaluator prompt scored 68.88%, below a 77.78% majority-class baseline; optimization raised it to 88.89%. The checker needed testing too. The debt-management agent improved but remained 23.6 transactional net-promoter-score points below expert humans. Case mix and incomplete comparison details limit causal interpretation. [9]

Brex repairs incomplete AI cost attribution

Published 1 June 2026 | Firsthand FinOps account

Brex’s AI-spend dashboards contained stale pricing assumptions and missing feeds; some vendors supplied only invoices, while gateway counts omitted tool charges. Phil Burrows described having responsibility for numbers that colleagues could not trust. Better attribution helped identify lower unit costs. He also said operational AI was cheaper than expected, with quality and integration more important to value than token prices. [15]

Brex captures investigation knowledge for an oncall agent

Published 27 April 2026 | Selected engineering and oncall workflow

Brex’s accounting-incident agent could reach logs and tables but lacked investigation strategies scattered across engineers, Slack and runbooks. The team found that improving runbooks mattered more than upgrading the model. It reported reducing context gathering for one export investigation from 30–45 minutes to about three. This is a selected engineering/on-call case, not a benchmark for all finance staff. [6]

Earlier fintech integration and repair accounts

Chime prepares the knowledge behind support deployment

Published 14 January 2026 | Company account

Chime describes restructuring content and sources of truth as part of successful support deployment. This shows the groundwork and maintenance behind usable assistance: procedures and interfaces still need owners when products, feeds and policies change. The account does not establish that every firm must replace its systems. [19]

Coinbase standardizes the production agent platform

Published 22 December 2025 | Company reported measurements

Coinbase found that low-code prototypes became difficult to reproduce and test as tools and instructions accumulated. Its production effort standardized hosting, evaluation and audit artifacts. The company reports cutting subsequent agent builds from more than twelve weeks to under one, and two automations saving more than 25 hours weekly. The initial platform burden can therefore be amortized rather than repeated in full. [16]

Klarna combines quality concerns with continuing AI use

SEC comment letter 30 May 2025 | Later FY 2025 annual report

Klarna illustrates why one headline is inadequate. SEC staff requested disclosure changes after reported service-quality concerns in 2025. Klarna’s later annual report still reported 80% of chats handled by AI, estimated $59 million in savings and a human option. Different periods and measures prevent a simple success-or-failure verdict. [17][18]

The SEC letter requested disclosure changes; it was not a finding that all AI service failed. Chat usage, savings and satisfaction in the annual report are company estimates or internal measures. [17][18]

Finom shortens the wait between finding and fixing agent problems

Account undated | Vendor relationship begins January 2026 | Not assigned a September event date

Finom’s AI-quality leader described prompt, test and metric changes queued behind engineering, plus manual or missing logs and inadequate conversational tests. A validated fix took a reported three to ten days despite roughly half a day of engineering work. [56]

Shared evaluation tools and realistic multi-turn tests reportedly cut the cycle to one to three hours. This is a materially improved handoff bottleneck. The case says agents executing real financial transactions were preparing to launch; it does not establish those actions were already live. Projected engineering savings are not realized cash savings. Vendor reporting supplies no production error rate. [56]

Earlier organizational and workforce accounts

Worldline employee representatives question access and workload

Meeting 22 May 2026 | Report published 12 June 2026 | Employee-representative account

CFTC representatives reported questions about unequal AI access, evidence of value and cognitive workload. Their account says Claude Code requests had to show business impact, such as staffing, revenue or service effects. Obtaining access therefore required a business case while workforce consequences remained contested. [57]

The reported management response favored flexibility across vendors and models. This is a governance and labor disagreement, not proven AI harm. The union account is not neutral meeting minutes, and it does not quantify changed working conditions or verify every management statement independently. [57]

Brex changes service roles and procurement

Published 25 September 2025 | Firsthand COO and CTO accounts

Brex’s COO describes more than fifteen service roles moving into a higher-complexity tier, changes to outsourced support and new expectations for managers to analyze and improve agents. Its CTO describes replacing slow tool-by-tool procurement with approval categories. These are changes in role design and operating practice, not evidence of net job losses or a representative staffing ratio. [21]

ING redesigns the approval path for deployment

Published 1 July 2025 | COO and transformation chief interview

ING’s COO said its existing risk-assessment process was not ready for generative AI. The bank introduced a senior-led review route, narrowed its initial domains and freed implementation staff from other duties. It reported reaching the first initiative in six or seven months. This is a case of controls enabling deployment after redesign, rather than a failed rollout. [30]

Zurich considers how future underwriters will learn

Published 29 September 2025 | Forward looking CEO concern

Zurich’s CEO raises a further concern: how to train underwriters when automation removes foundational tasks. This is a forward-looking operator concern, not measured skill deterioration. Repeated promotion and role-specific training may improve uptake, but they do not settle how expertise, authority or incentives should change. [27]

Earlier insurance workflow and evidence accounts

GEICO changes document requests and cancellation notices

Agreement announced 22 May 2026 | Underlying policyholder event undated

Pennsylvania’s Attorney General described an AI-featured tool selecting a policyholder for underwriting review. She believed she had supplied evidence, was not clearly told it was inadequate and unknowingly drove uninsured after cancellation. The strongest supported problem is the complete review-and-notice process. [55]

GEICO agreed to more time, simpler residency evidence and staff training, without admitting a violation. The announcement does not establish model misclassification, discrimination or autonomous cancellation. No post-remediation outcome data are reported. This is an official operational agreement, not a demonstrated AI selection-error rate. [55]

Zurich distinguishes a weak pilot from useful summaries

Recorded 16 October 2025 | Published 3 December 2025

An early loss-description classification pilot at Zurich reached about 60% accuracy after several weeks. The account gives no test-set size or human baseline and does not describe a production failure. The same interview reports useful submission summarization. [8]

Allianz links integration work to acquired systems

Published 29 January 2025 | Operations and data leader interview

Allianz’s operations and data leader links AI scaling work to diverse systems inherited through acquisitions. Commercial submissions and claims contain narrative evidence requiring specialist interpretation; data across policy and claims systems complicates reuse. This establishes preparation and integration work rather than failure caused by AI. [28]

Admiral and Aviva report manipulated claims evidence

Admiral 7 April 2026 | Aviva 8 June 2026 using 2025 internal data

Admiral describes manipulated claim images and documents, including changed vehicle registrations. Its reported 71% increase in the value of all detected fraud does not isolate AI and is affected by detection improvements. Aviva likewise reports more AI-generated images and altered documents; its headline fraud total also changed scope by including Direct Line brands. Neither total measures AI-attributable net loss. [22][23]

FinCEN flags deepfake identity material

Alert published November 2024 | Suspicious activity reporting

FinCEN’s 2024 alert describes increased suspicious-activity reporting involving deepfake media, including identity documents. A suspicious-activity report is not a confirmed fraud conviction or a measured failure of the institution’s own AI. [24]

Earlier lending and borrower workflows

AALTO corrects borrower generated misinformation

Practice described in 25 June 2026 publication | Interview date undisclosed

Stuart Phillips reported correcting invented or non-UK mortgage terminology that borrowers brought from AI tools. Matching several criteria was also difficult when lender wording was ambiguous or dispersed across documents. The work reached the broker even without the broker deploying the originating assistant. [52]

The consequence was explanation and expectation-setting work. Human verification remained necessary while Phillips continued experimenting with AI for cross-referencing criteria. The account supplies no error frequency or rework-time measure. His hypothetical model-cost estimate is not evidence of actual spending and is excluded here. [52]

Target Group assembles context for arrears advisers

In use by June 2026 | First results 2 June 2026 | Follow-up publication 3 September 2026

Target Group described arrears advisers switching among case histories, detailed procedures and multiple systems. Collections Guru assembled context and proposed next steps for human advisers. June reporting claimed a 30 percent reduction in average handling time, without a sample, measurement window or controlled comparison. [53]

This is bounded augmentation of a sensitive workflow. The September update is not a new rollout date; it also describes a 40 percent efficiency figure inconsistently as delivered and anticipated. That figure is not treated as established savings. The sources do not show improved recovery outcomes or the elimination of compliance risk. [54]

Upstart discloses credit cohort underperformance

FY 2025 filing dated 10 February 2026 | Earlier cohorts and later forecasts

Upstart disclosed that its models did not predict severe post-pandemic macroeconomic changes, with high defaults and charge-offs affecting fair values and revenue. Its FY2025 filing still forecast some late-2023 and early-2024 cohorts below target returns, while forecasting newer cohorts in line with targets. This is real disclosed credit-performance difficulty. It does not show AI alone caused it or that conventional models would have performed better. [14]

Marine Credit Union retains lending exceptions

Published 18 November 2025 | Management reported implementation history

Marine Credit Union expected 60–90 days of implementation but reported ten months to get through implementation and user acceptance testing and six months of postlaunch review. It settled near 55–60% automated decisions, below an 80% target, because going higher would exclude borrowers humans could help. These management-reported results show an intentional boundary, not a universal automation ceiling. [10]

Longbridge uses bounded automation for loan boarding

Published 15 July 2025 | Company reported workflow results

Longbridge’s loan boarding required staff to compare origination records, servicing records and documents. Its bounded OCR and machine-learning exception process reportedly cut cost per boarded loan nearly 50% and ACH errors 67%; prior draw errors could generate late fees. This is evidence of old workflow pain and a successful response, not AI-caused reconciliation loss. [29]

Surveys and regulatory context

These sources corroborate concerns, practices and adoption barriers within their respondents. They are not incident counts or representative estimates of project failure. Survey fieldwork dates and question denominators matter.

Cambridge financial services survey

Published April 2026 | Fieldwork October 2025 to January 2026

Cambridge surveyed 352 financial firms, including 203 fintechs and 149 traditional institutions. Among 344 barrier-question respondents, talent (41%) and data (40%) led resources (29%), infrastructure (28%), value tracking (24%) and regulation (23%). Recruitment may favor AI-engaged firms. These are reported barriers, not project failure rates. [1]

London market and European insurer surveys

LMA published 24 April 2025 | EIOPA published 2 February 2026

In the Lloyd’s Market Association’s survey of 81 firms, including 45 managing agents, 49% cited data availability or quality, 46% integration and 48% cost or uncertain return as barriers. The public release does not provide a response rate or field dates. [4]

EIOPA’s survey of 347 insurers in 25 EU and EEA countries identifies privacy/security, compliance and skills among leading implementation barriers. Fieldwork ran May–July 2025 and adoption status relates to Q2 2025. Practical work includes deciding what customer data providers receive, who may access logs and who remains responsible for consequential decisions. Stated concerns are not evidence of a breach. [5]

Fannie Mae provides a historical lender baseline

Survey August 2023 | Published October 2023

The survey covered 242 executives at 219 lenders. Among 162 nonusers answering the integration question, 48% put integration among their top two barriers. This was a concern before deployment, not a measured failure or a measure of current prevalence. [13]

NAIC documents continuing supplier discovery

Report May 2025 | 93 selected insurers under 16 states’ market conduct authority

A respondent to NAIC’s health-insurer survey described repeated vendor surveys to discover AI added after procurement, through acquisitions or previously missed. The report covers 93 selected insurers; the same response repeated under different vendor categories is not multiple independent examples. [20]

Regulatory review and anonymous recent accounts

Bank of England roundtables describe review delays

Three late 2025 roundtables | Summary published 16 February 2026

At the Bank of England’s late-2025 roundtables, banks and insurers described second-line review delays, supplier contracting difficulties and cross-border data requirements. Most supported the existing principles-based PRA framework. Views differed on whether caution was excessive. The anonymous summary gives no participant denominator or measured average delay. [2]

UK surveys and consultations describe assurance work

2024 survey | September 2025 feedback | June 2026 workshop

In the Bank of England/FCA survey of 118 responding firms, 46% of AI users or planned users reported only partial understanding of their technology, particularly third-party models. The measure covers broad AI, not solely generative AI. [3]

The FCA’s AI Live Testing consultation recorded difficulty defining acceptable performance and translating lab tests into live behavior; only 15 of its 67 respondents were regulated firms. [25] The AI Consortium’s 3 June 2026 meeting, published 5 August, discussed independently changing components, oversight and limited transparency. Systemic outage examples were scenarios, not a documented catastrophe. [26]

Anonymous recent evidence

At a 20 to 22 September mortgage-compliance session, an unnamed lender reportedly described an unannounced vendor switch to AI quality review. Monthly findings rose from 3-5 per month to more than 30, with inconsistent severity labels and two months of calls. The underlying vendor-change date is unknown. The 30 September recap gives no file denominator or ground truth; more findings alone could mean better detection. The affected lender is not identified. [63]

Moody’s 29 September study describes incomplete inventories and uneven governance evidence among some of five interviewed banking groups. Decision authority generally remained human; autonomous action was mostly discussed or piloted. Interview dates within 2026 are undisclosed, and the sample is not representative. These are reported practices, not five failure cases. [64]

LenderBox’s 7 September account of ninety institutions interviewed over nineteen months found weak baseline knowledge and vendor fatigue. Its combined category of vendor fatigue or failed implementation cannot be counted as failed AI projects. Participants were anonymous, commercially recruited and geographically concentrated. [65]

Technical tests and measured counterevidence

Research experiments reveal mechanisms but are not additional production failures. The newest result matters when earlier benchmarks are used to describe current capability.

Accounting results changed with the task and comparison

Latest update 1 October 2026 | Earlier benchmark 29 July 2026

Mercor’s October comparison had frontier models score perfectly on four simplified accounting tasks, while twelve licensed CPAs averaged about 37 percent. Participants lacked coworkers and accumulated company context. The small, constrained task set is not a test of complete professional workflows. [58]

The July APEX study used 160 deliberately difficult synthetic tasks. Its analyzed failures were mainly accounting reasoning, rather than tool use. That result and the October result use different task sets and comparisons; neither is an industry failure rate or evidence of a real incorrect transfer. The evaluator and commercial research incentives also warrant scrutiny. [59][58]

NPCI checks whether banking actions fit the customer state

Paper published 3 September 2026 | Synthetic testing rather than live customer incidents

NPCI found plausible tool arguments unsupported by customer state. Training could also exploit stylistic differences in preferred answers. Its response combined state, sequence and communicated-value checks with a limited model judge. Improvements in evaluation do not establish live-bank outcomes or isolate every training choice’s contribution. [60]

Ramp finds that memory can preserve the wrong accounting fact

Account published 3 June 2026 | Experiment dates undisclosed

In one synthetic next-period billing task, remembered prior-period instructions cut a score from 75 to 25 percent. Ramp shifted toward procedural memory and filtering period-specific facts. Across its roll-forward evaluation, mean scores instead improved from 56.0 to 61.5 percent. Memory had conditional value; this was not a customer’s erroneous journal entry or measured lost money. [61]

Lemonade reports an actual handling cost measure

Observed April to June 2026 | Published 29 July 2026

Lemonade reported a 5 percent claims-handling expense ratio, versus 7 percent at an earlier premium milestone, attributing efficiency partly to AI. This is company-disclosed operating counterevidence. The ratio uses gross earned premium and does not isolate AI from scale, product mix or other process changes. [62]

Interpretation and remaining uncertainty

What the operating cases establish

A correct step can still produce a poor overall result. An answer, extracted field, generated memo or faster code change has to reach the right person, system and decision. Measures of complete, correctly resolved cases are more informative than isolated output speed.

Review requires a maintainable operating process. Someone must own changing policies, permission boundaries, exceptions and the path from an observed error to a validated fix. Retained review can be an intentional control, but its labor and effectiveness still need measurement.

Customer and employee behavior creates additional work. Borrowers may arrive with AI-generated misconceptions, employees may adopt tools before shared controls exist, and providers may change methods. These are different causes and should not be assigned to a single model-quality category.

Where the sectors differ

Fintech and payments. Customer-service and financial-crime workflows emphasize correct action, system state, fast escalation and variable transaction volumes. Published technical cases expose build and QA work, but overrepresent well-resourced firms that can describe their own platforms. These are research patterns, not measured differences in failure rates.

Insurance and insurtech. Commercial submissions and claims contain narrative evidence requiring specialist interpretation. Data accumulated across policy and claims systems complicates reuse. External evidence manipulation is especially visible. Evidence is stronger for established carriers than small insurtechs. [28]

Lending. Individual credit exceptions and delayed repayment outcomes create different problems from document checking at loan boarding. The distinction matters when interpreting automation percentages: workflows have different objectives, consequences and observation periods. [13][29]

Limits of the conclusions

This is a synthesis of public evidence available by 2 October 2026, not a representative operator interview study. US, UK and European institutions dominate. Executive and engineering accounts are better represented than frontline staff, small firms, abandoned projects or unpublished incidents. A solved historical case demonstrates a mechanism, not necessarily a current unresolved need.

Across the cases, public evidence rarely supplies net reviewer hours, fully loaded cost per correctly resolved case, failure frequency, customer harm, post-change recurrence or an independent prior-workflow comparison. Named accounts are often promotional; anonymous accounts reduce verifiability. Several workflows at one institution are not independent replications. The evidence supports specific mechanisms and workarounds, not sector-wide failure estimates or willingness to pay.

Sources

Numbered citations link directly to their source. Dates are publication dates unless otherwise stated; event and observation dates accompany the cases. Attributed company metrics, survey limitations and source-specific caveats remain in the body and entries below.

1. Cambridge Centre for Alternative Finance and Financial Innovation for Impact. Global AI in Financial Services Report 2026 (April 2026). Voluntary survey; methods pp 14–15, value pp 67–68, barriers p 74. Fieldwork October 2025 to January 2026.

2. Bank of England. Summary of AI roundtables (16 February 2026). Three late 2025 banking and insurance roundtables; anonymized operator views, no participant denominator.

3. Bank of England and Financial Conduct Authority. Artificial intelligence in UK financial services 2024 (21 November 2024). 118 responding firms; self-reported adoption, model understanding and constraints. Broad AI, not solely generative AI.

4. Lloyd’s Market Association. AI adoption in underwriting and claims survey (24 April 2025). 81 London-market firms, including 45 managing agents; public release does not provide response rate or field dates.

5. European Insurance and Occupational Pensions Authority. Generative AI market survey outlook use cases and risk management (2 February 2026). 347 insurers in 25 EU and EEA countries, surveyed May–July 2025; adoption status relates to Q2 2025.

6. Jiwoo Hong at Brex. How we built an AI oncall engineer (27 April 2026). Firsthand engineering account; internal evaluations and selected incident timings.

7. Jamie McDonald-Gibson and colleagues at Monzo. Engineering the future of customer operations the Monzo Ops Agent (June 2026). Firsthand production engineering account. Live page and index differ by one publication day.

8. Amy Nelsen and Madhu Ramamurthy at Zurich North America. What is next in AI (3 December 2025). Insurer podcast recorded 16 October 2025; firsthand pilot and successful deployment accounts, without test-set detail.

9. Aman Gupta and colleagues at Nubank. Building Customer Support AI Agents at 100M-User Scale: An Evaluation-Driven Framework (7 June 2026). KDD 2026 paper. Evaluator experiments in sections 6.1–6.2; customer metrics in Table 3; privacy controls in 6.5.

10. Ken Brossman at Marine Credit Union interviewed by PixelSpoke. Beyond the hype inside a credit union’s AI lending journey (18 November 2025). Firsthand chief lending officer account; implementation, review and automation figures are management-reported.

11. Sarah Gonzalez at Logan Finance. Eliminate the task not the judgment AI in non QM lending (22 September 2026). Firsthand CIO interview hosted by technology vendors; benefits and implementation-cost criticism are attributed claims.

12. Lindsey Ricciardelli at American Pacific Mortgage. Stop training start marketing AI adoption in mortgage lending (15 September 2026). Firsthand senior-operator interview hosted by technology vendors; no measured adoption rate supplied.

13. Fannie Mae. Mortgage Lender Sentiment Survey special topic artificial intelligence (October 2023). August 2023 survey: 242 executives at 219 lenders; nonuser integration question n 162. Historical baseline, not current prevalence.

14. Upstart Holdings. Annual report 2025 Form 10 K (10 February 2026). Credit Performance and Risk Factors disclose cohort underperformance and forecast improvement. AI-only causation is not established.

15. Phil Burrows at Brex. Magpie AI cost visibility dashboard (1 June 2026). Firsthand FinOps account of incomplete cost feeds, attribution and unit-cost changes.

16. Jason Dodds at Coinbase. Building enterprise AI agents at Coinbase (22 December 2025). Firsthand platform account; build-time and hours-saved figures are company-reported.

17. US Securities and Exchange Commission staff. Klarna registration statement comment letter (30 May 2025). Comment 2 cites reported service-quality concerns and requests revised disclosure; not a finding that all AI service failed.

18. Klarna Group. Annual report 2025 Form 20 F (FY 2025 report). PDF pp 87 and 188. Chat usage, savings and satisfaction are company estimates or internal measures.

19. Chime with COO Janelle Sallenave. How Chime is using AI to redefine member experience (14 January 2026). Company account of knowledge preparation, support outcomes and dispute handling; metric definitions are limited.

20. National Association of Insurance Commissioners. Health insurance AI and machine learning survey report (May 2025). 93 insurers selected under 16 states’ market-conduct authority. Method p 10; vendor discovery responses pp 208–211.

21. Camilla Matias and James Reggio at Brex interviewed by First Round. How Brex is building an AI first company (25 September 2025). Firsthand COO and CTO accounts of role changes, procurement and tool adoption; selected company case.

22. Admiral Group. The white lie effect and insurance claims fraud (7 April 2026). Carrier observations of manipulated claim evidence; overall detected fraud values do not isolate AI.

23. Aviva. Aviva stops record levels of claims fraud as scams become more sophisticated (8 June 2026). Carrier observations using 2025 internal data. Headline totals include Direct Line brands for the first time.

24. Financial Crimes Enforcement Network. Alert on fraud schemes involving deepfake media targeting financial institutions (November 2024). Government alert citing suspicious activity reporting; does not quantify confirmed AI-related loss incidence.

25. Financial Conduct Authority. AI Live Testing feedback statement FS25 5 (September 2025). 67 submissions, of which 15 were regulated firms; evaluation, integration and monitoring views on pp 5 and 7–12.

26. Bank of England and FCA AI Consortium. Minutes of the 3 June 2026 meeting (Published 5 August 2026). Expert workshop discussion of changing components, oversight and concentration; scenarios are not observed incidents.

27. Kristof Terryn at Zurich North America. From car dreams to corner office (29 September 2025). CEO discussion of training underwriters as foundational tasks become automated; forward-looking concern.

28. Manuela Diviach at Allianz. Scaling AI to transform insurance (29 January 2025). Firsthand operations and data interview linking diverse acquired systems to integration work and describing knowledge assistance.

29. Bill Packer at Longbridge Financial interviewed by HousingWire. How technology is influencing Longbridge workflows (15 July 2025). Firsthand COO account; boarding cost and ACH-error improvements are company-reported.

30. Marnix van Stiphout at ING interviewed by McKinsey. Deploying AI at speed and scale (1 July 2025). Firsthand COO and transformation chief account of risk review changes, dedicated staff and staged scaling.

31. Libra Internet Bank. Public statement about Lia (September 2026). Company acknowledgment and announced restrictions

32. HotNews. Report on Libra Internet Bank assistant and bank response (19 September 2026). Screenshot and company response linked

33. Timothy Wong at Airwallex. How Airwallex builds fintech AI agents (15 September 2026 original; 19 September republication). Current executive account

34. Konstantin Dzhengozov at Payhawk. CFO perspective on AI in finance (21 August 2026; updated 24 September). Supplier observations; affected customers unnamed

35. Sreevishnu Sachidananda at Wise. Moving upstream (28 July 2026). Current quality-operations interview

36. Tyran Elgar at Checkout.com. Operationalizing AI in engineering and beyond (22 July 2026). H1 and June measurements; current transformation account

37. Daniel Linder at Checkout.com. How we built an AI payment optimization system (13 July 2026). Current production account

38. NN Group and Tjerrie Smit. The story behind NN AI journey so far (24 September 2026). Interview and deployment dates undisclosed

39. AWS and MRH Trowe. How MRH Trowe enabled secure self service AI agents in financial services (17 September 2026). Joint implementation account

40. Reece Adams at Markel and Irina Simpson at Gallagher Bassett. AI is changing workers comp and small commercial but how will insurers build judgement (15 September 2026). Sponsored webinar account; program dates unstated

41. Jacquelene Robinson. Declaration concerning Irys and motion citations (11 August 2026). Primary sworn declaration; pages 2 to 7

42. Reuters via The Insurer. Lawyer for State Farm fined over AI hallucination in Los Angeles lawsuit (September 2026). Sanction reported; signed final order not separately inspected

43. Willard T Lee at The Hanover. Augmenting insurance with AI (1 September 2026). Sponsored operator Q and A

44. Ying Zhang Aloysius Lim and Samuel Lim at Gen Re. AI enabled underwriting 2 0 (4 August 2026). Continuing issues across two years of implementation

45. Germania Mutual and Guidewire. Germania Mutual deploys Guidewire AI to put institutional knowledge at the fingertips of employees (16 July 2026). Joint deployment announcement

46. Hasani Jess at Aventum. Aventum Group ecosystem turning to bespoke AI to scale specialty lines (8 July 2026). Current interview; investment began in 2025

47. Lindsey DiGangi at Pennsylvania Lumbermens Mutual. Why this wood industry mutual will not throw AI at everything (8 July 2026). Current COO interview

48. Devesh Maheshwari at Lendi Group. Lendi Group taps agent swarm to reshape mortgage broking (8 September 2026). Current CTO interview

49. Abdel Khawatmi at PRMG. Adopt AI or take a compensation cut (8 September 2026). Recorded late August; see 37:01 to 42:18

50. Nimish Panchmatia at DBS. DBS holds off on letting AI agents run on their own as controls lag capability (28 July 2026). Current operator interview

51. David Cedrone at Rockland Federal Credit Union. How one credit union used AI to triple loan production without hiring anyone (27 July 2026). Current interview; outcome window undisclosed

52. Stuart Phillips at AALTO Mortgages. AI could transform mortgage broking but price tag leaves firms counting the cost (25 June 2026). Current broker interview

53. Target Group and Modern Lender. Target Group drives 40 percent increase in efficiency with Collections Guru (2 June 2026). Company reported workflow results

54. Tech Mahindra. Tech Mahindra launches AWS agentic process transformation centre (3 September 2026). Target update contains delivered versus anticipated ambiguity

55. Pennsylvania Attorney General. Agreement with GEICO on underwriting review and cancellation communications (22 May 2026). Agreement without admission of violation

56. Igor Kolodkin at Finom with Confident AI. Finom case study (Undated; customer since January 2026). Vendor-published customer account

57. CFTC Worldline employee representatives. Report of the Worldline CSE meeting of 22 May (12 June 2026). Union account of management discussion

58. Mercor. Human baselines for benchmarks (1 October 2026). Four simplified tasks; twelve licensed CPAs

59. Mercor and Ramp researchers. APEX Accounting benchmark paper (29 July 2026). 160 synthetic accounting tasks

60. NPCI AI Research Team. FiMI Banking (3 September 2026). Synthetic customers and executable tools

61. Ryan Stevens at Ramp. Stack benchmarking (3 June 2026). Simulated-company tests; actual customer loss not established

62. Lemonade. Second quarter 2026 shareholder letter (29 July 2026). April to June observations; pages 3 to 4

63. Melissa Grindel at ActiveComply. Uniquely TMC 2026 recap (30 September 2026). Session 20 to 22 September; anonymous participant

64. Alain Laurin and Emmanuel Nusimovici at Moody’s. Governance before autonomy (29 September 2026). Anonymized 2026 interviews with five banking groups

65. LenderBox. What ninety lenders told us about AI (7 September 2026). Commercially recruited discovery interviews over nineteen months